Vaicoa Core
Vaicoa Core is an approval layer for AI agents in IT operations. It decides who may do what, and records what happened. An agent proposes an action from a signed catalog; the core computes the plan and the policy outcome; a human approves on a phone, a Mac or the command line by signing with a device key; and the node that runs the action checks that signature, the signed team roster and the plan itself before it changes anything.
The three layers
Section titled “The three layers”| Layer | What it does | Built in |
|---|---|---|
Core (vaicoa-core) | identities and a signed roster, its own CA, the signed action catalog, policy (Cedar and CEL), approvals and quorums, the queue, the hash-chained audit, secrets | Go, one binary |
| Modules | answer “is this safe now?” and produce a plan (Podman environments, k3s clusters, Cluster API). They can only make the core stricter, never looser. | Python, out of process |
| Executors and drivers | do the work: vaicoa-node on each host (Podman, k3s, Kubernetes, Cluster API); drivers at the core for third-party APIs (Hetzner Cloud, Forgejo, Alertmanager, observability, MCP servers, a local model) | Go (node), Python (drivers) |
Clients: vcx (the command line for humans), vaicoa-run (the client for coding agents, a CLI and a local MCP
server), and the approval apps for iPhone, Android and macOS. Details in architecture.
Who it is for
Section titled “Who it is for”Small ops and platform teams, from one person to a few, who run their own infrastructure (a server with Podman, k3s, Cluster API, Hetzner Cloud) and want coding agents to do operations work without handing them the keys. It is also material for anyone interested in how “DevOps with agents” can be made controllable and traceable.
It is not for you if you need a managed service, a supported product, or something to run in production now.
Where to go next
Section titled “Where to go next”- Quickstart: your own core, alone, in a lab.
- Approvals: what the approver sees and signs.
- Security model and threat model.
- Limitations: what is known to be missing or unproven.
The user manual in the repository (docs/handbuch.md, German, about 190 pages) walks through everything by hand,
with expected output for every step. These pages are a short English digest of it.